Download O A

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Mathematical model wikipedia , lookup

Perceptual control theory wikipedia , lookup

Pattern recognition wikipedia , lookup

Gene expression programming wikipedia , lookup

Linear belief function wikipedia , lookup

Neural modeling fields wikipedia , lookup

Hierarchical temporal memory wikipedia , lookup

Catastrophic interference wikipedia , lookup

Convolutional neural network wikipedia , lookup

Transcript
5692
Journal of Applied Sciences Research, 9(11): 5692-5700, 2013
ISSN 1819-544X
This is a refereed journal and all articles are professionally screened and reviewed
ORIGINAL ARTICLES
A comparison of methods for assessing the relative importance of input variables in
artificial neural networks
O.M. Ibrahim
Field Crops Research Department, National Research Centre, Giza, Egypt.
ABSTRACT
Artificial neural networks are considering powerful statistical modeling technique in the agricultural
sciences; however, they provide little information about the contributions of the independent variables in the
prediction process. The goal of relative importance analysis is to partition explained variance among multiple
predictors to better understand the role played by each predictor. In the present study, a modification to
Connection Weights Algorithm and a novel algorithm are proposed to assess the relative importance of
independent variables in multilayer perceptron neural network and a comparison in the field of crop production
with the Connection Weights Algorithm, Dominance Analysis, Garson’s Algorithm, Partial Derivatives, and
Multiple Linear Regression is presented. The performance of the two proposed algorithms is studied for
empirical data. The Most Squares method (the second proposed algorithm) is found to be a better method in
comparison to the above mentioned methods and seem to perform much better than the other methods, and
agree with the results of multiple linear regressions in terms of the partial R2 and consequently, it seem to be
more reliable.
Key words: Artificial neural networks, relative importance, statistical modeling.
Introduction
The relationships between variables in agriculture are almost very complicated. One of the most appropriate
methods to illustrate these relationships is Artificial Neural Networks (ANNs). A number of authors have shown
the interest of using ANNs instead of linear statistical models (Özesmi and Özesmi, 1999). The main application
of ANNs is the development of predictive models to predict future values of a particular dependent variable
from a given set of independent variables. The contribution of the input variables in predicting the output is
difficult to be explained within the network. Consequently, input variables are entered into the network and an
output value is generated without understanding the inter-relationships between the variables, and therefore,
providing no explanation by the network (Ripley, 1996). The study of contributions of input variables for ANNs
models has been attempted only by few authors. For example (Duh, et al., 1998) have presented a methodology
to understand how an input is correlated to the predicted output by the ANNs. They have tested this
methodology on three datasets and have shown that the results correspond well to the partial least squares
method interpretation for linear models. (Olden and Jackson, 2002) have reviewed a number of methods (Neural
Interpretation Diagram, Garson’s algorithm and sensitivity analysis) and demonstrated utility of these methods
for interpreting neural network connection weights. (Olden et al., 2004) have provided a comparison of different
methodologies for assessing variable contributions in artificial neural networks using simulated data exhibiting
defined numeric relationships between a response variable and a set of predictor variables. They proposed an
approach called connection weights method to outperform other methods in quantifying importance of variables.
They have shown that the connection weight method is the least biased among others. This method has also
been used in comparison to other available methods in assigning the relative contribution of input variables in
prediction of the output by (Watts and Worner, 2008).
Materials and Methods
An experiment was set up during the summer seasons of 2012 and 2013, Soil Salinity Laboratory,
Alexandria, Egypt to investigate the effect of irrigation with saline water (0.5, 2.75, 5.5 dS/m) on yield and yield
components of maize as well as the relative importance of yield components. Well water was used as a source of
saline water. Grains of maize cultivar (Gemmeiza 12) were obtained from Crops Research Institute, Agricultural
Research Center, Egypt. The soil and irrigation water were analyzed according to (Chapman and Pratt, 1978)
before sowing and had the following characteristics:
Corresponding Author: Ibrahim, O.M., Field Crops Res. Department, National Research Centre, Giza, Egypt.
E-mail: [email protected]
5693
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
Table 1: Physical and chemical analyses of soil before sowing.
Sand %
Silt %
Clay %
Soil texture
74.0
10.4
15.6
Sandy loam
Table 2: Chemical analyses of irrigation water.
Treatments
PH
E.C.
dS/m
Tap water
7.6
0.62
Well water
7.8
5.44
Ca2+
2.2
1.8
E.C. dS/m
1.82
pH
7.53
Cations, meq/ L.
Mg2+
K+
1.73
0.18
2.4
3.9
Na+
2.75
48.0
S.P. %
43.33
HCO33.55
2.9
SAR
1.49
Anions, meq/L.
ClSO422.11
1.2
32.0
13.9
The experimental design was Randomized Complete Block Design (RCBD) with four replicates; the
experimental unit was a cemented plot with a dimension of 150 cm in long and 75 cm in wide with an area of
1.125 m2. Every cemented plot contains four rows, the grain were sown in the first of May and before sowing
the plots were prepared by adding calcium superphosphate 15.5% P2O5 at a rate of 100 kg/feddan (Hectare =
2.38 feddan) and potassium sulphate 48% K2O4 at a rate of 50 kg/feddan, the nitrogen fertilizer rates were added
at the rate of 125 kg N/feddan as ammonium sulphate 20.5% N at three doses, the first at sowing, the second at
the first irrigation, and the third at the second irrigation.
At the end of the experiment the following characters were measured:
Number of rows/ear, number of grains/row, 100 kernels weight (g), and grain yield (g/plot).
The data were subjected to the analysis of standardized and stepwise multiple linear regression according to
(Snedecor and Cochran, 1982) and dominance analysis using SAS computer software (SAS Institute Inc., 2003).
A program was proposed by the author, based on Microsoft Access 2007 and containing modules written in
visual basic computer language, was used to perform the comparison among the different relative importance
methods, figure 2.
Neural network training and architecture:
A three layer feed forward neural network was trained using back propagation algorithm. The three layered
feed forward network contains one input layer, one hidden layer and one output layer. The input layer contains 3
neurons corresponding to 3 independent variables, number of rows/ear (X1), number of grains/row (X2), and
100 kernels weight (g) (X3) whereas the output layer contains one neuron corresponding to one dependent
variable, grain yield (g/plot) (GY). The sigmoid activation function was used at both the hidden layer and the
output layer. The structure of the multilayer perceptron neural network used in this study and its connections
between layers is shown in Figure 1.
Fig. 1: Structure of the multilayer perceptron neural network used in this study.
5694
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
Fig. 2: Interface of the proposed program used to perform the comparison among the studied methods.
Results and Discussion
Effect of water salinity on yield and yield components of maize:
Data presented in Table (3) show that irrigation with saline water adversely and significantly influenced
number of grains/row, 100 kernels weight (g), and grain yield of maize plants (g/plot) as compared with control
plants. However, number of rows/ear was not significantly affected by water salinity. The reduction in grain
yield of maize plants was 32.13% and 61.31% when the plants were irrigated with 2.5 and 5.5 dS/m,
respectively. These results are in agreement with (Khodary, 2004).
Table 3: Effect of water salinity on number of rows/ear, number of grains/row, 100 kernels weight (g), and grain yield (g/plot) of maize
plants, (combined analysis of the two seasons).
Number of Rows/ear
Number of
100 kernels weight(g)
Grain yield(g/plot)
Salinity
grains/row
Control
11.50 a
30.25 a
23.68 a
407.02 a
2.5 dS/m
11.50 a
19.25 b
20.40 a
262.03 b
5.5 dS/m
12.00 a
15.75 b
17.25 b
170.50 c
Relative importance of independent variables methods:
The relative importance of predictor or input variables is the contribution of each of the variables for
predicting the dependent variable. The proposed methods are defined after a brief description of the connection
weights algorithm in order to quantify the relative importance of independent variables in predicting the output
variable for neural network.
1- Connection weights algorithm (CW):
The connection weights algorithm (Olden and Jackson, 2002) calculates the sum of products, Table (5) of
final weights of the connections from input neurons to hidden neurons, Table (4) with the connections from
hidden neurons to output neuron, Table (4) for all input neurons. The relative importance of a given input
variable can be defined as:
m
RI x = ∑ wxy w yz
y =1
m
Where
RI x is the relative importance of input neuron x ,
∑w
y =1
xy
wyz is sum of product of final weights of
the connection from input neuron to hidden neurons with the connection from hidden neurons to output neuron,
5695
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
y is the total number of hidden neurons, and z is output neurons. This approach is based on estimates of network
final weights obtained by training the network. It is observed that these estimates of final weights may vary with
the change in the initial weights used for starting the training process.
Table 4: Final connection weights.
X1
X2
X3
Hidden1
-0.8604
-6.6167
-1.7338
GY
Hidden1
-5.2075
Input-hidden
Hidden2
1.0651
1.5145
1.4107
Hidden-output
Hidden2
2.0158
Hidden3
-0.1392
-5.3051
-1.4207
Hidden4
1.0666
1.3358
0.7845
Hidden3
-4.4617
Hidden4
1.5858
Table 5: Connection weights products, relative importance and rank of inputs.
Hidden1
Hidden2
Hidden3
Hidden4
X1
X2
X3
Sum
4.4805
34.4562
9.0286
47.97
2.1469
3.0529
2.8436
8.04
0.6212
23.6698
6.3388
30.63
Sum
1.6914
2.1184
1.2442
5.05
8.94
63.30
19.46
91.69
Relative
importance
9.75%
69.03%
21.22%
100%
Rank
3
1
2
2- The first proposed algorithm (MCW):
This method is referred to as Modified Connection Weights. In this proposed algorithm, the connection
weights of the artificial neural network model is obtained after training the network, after the calculation of sum
of product of final weights of the connections from input neurons to hidden neurons with the connections from
hidden neurons to output neuron for all input neurons, a correction term (partial correlation) is multiplied by this
sum and the absolute value is taken, this is called the corrected sum, Table (6), then to calculate the relative
importance of each input, the corrected sum of each input is divided by the total corrected sum as illustrated in
the following equation.
m
RI x =
∑w
y =1
xy
wyz × rij .k
n
m
x =1
y =1
∑ ∑w
xy
wyz
m
where
RI x is the relative importance of neuron x,
∑w
y =1
xy
wyz is sum of product of raw weights of the
connection from input neuron to hidden neurons with the connection from hidden neurons to output neuron,
is partial correlation of input i with output j after input k,
n
m
x =1
y =1
∑ ∑w
xy
rij .k
wyz is the total corrected sum of all
inputs.
The correction term is the partial correlation; the partial correlation is the correlation that remains between
two variables after removing the correlation that is due to their mutual association with the other variables. The
correlation between the dependent variable and an independent variable when the linear effects of the other
independent variables in the model have been removed.
The partial correlation for first order is illustrated in the following equation:
rij .k =
where
rij − rki × rkj
(1 − r )× (1 − r )
2
ki
2
kj
rij .k is partial correlation of input i with output j after input k, rij is the simple correlation between input
i and output j, rki is the simple correlation between input k and input i,
input k and output j.
rkj is the simple correlation between
5696
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
Table 6: Absolute corrected sum and relative importance of inputs.
Sum
partial correlation
Inputs
X1
8.94
0.5266
X2
63.30
0.9584
X3
19.46
0.7275
Total
91.69
Absolute corrected
Sum
4.71
60.66
14.15
79.53
Relative importance
Rank
5.920%
76.28%
17.80%
100%
3
1
2
3-The second proposed algorithm (MS):
This method is referred to as Most Squares. In this proposed algorithm, the connection weights between
hidden layer and the output layer were not used; however, the connection weights between input layer and
hidden layer were used, both the initial weights before start training, Table (7) and the final weights after
training the network, Table (4). The second step is to sum the squared difference between initial and final
weights for each input. The third step is to divide the sum of squared difference for each input on the total sum
of all inputs, Table (8). The following equation is used to calculate the relative importance of each input.
∑ (w
m
RI x =
i
xy
x =1
f
− wxy
∑ ∑ (w
m
x =1
n
y =1
i
xy
)
2
f
− wxy
)
2
∑ (w
m
RI x is the relative importance of neuron x,
where
x =1
i
xy
− wxyf
)
2
is the sum squared difference between
initial connection weights and final connection weights from input layer to hidden layer, and
∑ ∑ (w
m
x =1
n
y =1
i
xy
− wxyf
)
2
is the total of sum squared difference of all inputs.
Table 7: Input-hidden initial connection weights.
Hidden1
X1
0.0236
X2
0.0562
X3
-0.1488
Hidden2
0.1943
-0.3437
0.1844
Hidden3
0.1478
-0.3215
-0.0608
Table 8: Squared difference between initial and final connection weights, relative importance, and rank of inputs.
Hidden1
Hidden2
Hidden3
Hidden4
Relative
Sum
importance
X1
0.7815
0.7582
0.0824
0.7416
2.3638
2.84%
X2
44.5279
3.4528
24.8357
2.0313
74.8478
89.87%
X3
2.5121
1.5039
1.8493
0.2096
6.0749
7.29%
Sum
83.2865
100.00%
Hidden4
0.2054
-0.0894
0.3267
Rank
3
1
2
4- Multiple linear regression (MLR):
A comparison between MLR and the other methods was undertaken in order to judge their predictive
capabilities. The stepwise multiple regression technique (Weisberg, 1980 and Tomassone et al., 1983) was
computed especially to define the significant variables and their contribution order. In fact the influence of each
variable can be roughly assessed by checking the final values of the regression coefficients. Standardized
regression coefficient has been suggested as a measure of relative importance of input variables by many
investigators (Afifi and Clarke, 1990). For each input variable, standardized regression coefficient is obtained by
standardizing the variable to zero mean and unit standard deviation before multiple linear regression is carried
out. However, when input variables are correlated, the influence of correlations between input variables makes
standardized regression coefficient not interpretable in explaining the relative importance (Johnson, 2000).
Researchers have developed other indices that accurately reflect the contribution of input variables to the
prediction of a dependent variable when variables are correlated. The two most recent methodologies are
dominance analysis (Budescu, 1993, and Azen and Budescu, 2003) and Johnson’s epsilon (also referred to as
relative weights; (Johnson, 2000). The results of multiple linear regression are shown in Table (9). The results
revealed that number of grains/row (X2) was the most important variable with a partial R2 of 0.8568 followed
by 100 kernels weight (g) (X3) (0.0552), and number of rows/ear (X1) (0.0217). To make a comparison with the
other methods, the relative importance of each input was computed by dividing its partial R2 on the sum of R2
for all inputs which is equal to regression R2 (0.9337), for example the relative importance of number of
grains/row (X2) is computed as: RI = (0.8568/0.9337)*100 = 91.76 %.
5697
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
5- Dominance analysis (DA):
DA determines the dominance of one input over another by comparing their additional contributions across
all subset models. Dominance analysis is chosen over other recent measures for reasons stated by (Lebreton et
al., 2004). Dominance analysis (Budescu, 1993) approaches the problem of relative importance by examining
the change in R2 resulting from adding an input to all possible subset regression models. By averaging all of the
possible models (average squared semi partial correlation), we can obtain general dominance weight of an input,
which reflects the contribution by itself and in combination with the other inputs, and overcoming the problems
associated with correlated inputs. The results of dominance analysis are presented in Table (10). To make a
comparison with the other methods, the relative importance of each input was computed by dividing its overall
average contributions on the sum of all overall average contributions for all inputs which is equal to regression
R2 (0.9337), for example the relative importance (RI) of number of grains/row (X2) is computed as follows: RI
= (0.7836/0.9337)*100 = 83.92%.
Table 9: Results of multiple linear regression and correlation.
Parameter
Standardized
estimates
estimates
Significance
Intercept
-547.161
X1
21.185
0.1514
0.0001
X2
21.270
0.8622
0.0001
X3
10.452
0.2764
0.0001
Partial R2
0.0217
0.8568
0.0552
Table 10: Relative importance and rank of inputs.
Overall average contributions of inputs
Inputs
X1
0.0198
X2
0.7836
X3
0.1303
Total
Relative importance
based on contribution
to regression R2
2.32%
91.76%
5.91%
Simple
Correlation
Partial
Correlation
0.1088
0.9256
0.4310
0.4965
0.9560
0.7138
Relative importance based on contribution to
regression R2
2.12%
83.92%
13.96%
100.00%
Rank
3
1
2
6- Garson’s algorithm (GA):
(Garson, 1991) proposed a method for partitioning the neural network connection weights in order to
determine the relative importance of each input variable in the network. It is important to note that Garson’s
algorithm uses the absolute values of the final connection weights when calculating variable contributions, and
therefore does not provide the direction of the relationship between the input and output variables as illustrated
in the following equation.
wxy w yz
n
RI x = ∑
x =1
m
∑w
y =1
xy
w yz
m
where
RI x is the relative importance of neuron x,
∑w
y =1
xy
wyz is sum of product of final weights of the
connections from input neurons to hidden neurons with the connections from hidden neurons to output neurons.
Table (11) shows the contribution of each input to each hidden neuron, for example the contribution of number
of grains/row (X2) is (0.7184) resulted from dividing its products (34.4562) on the sum of products of all inputs
(47.97) in Table (5) and take the absolute result.
Table 11: Relative contribution of each input neuron to each hidden neuron, relative importance, and rank of inputs.
Hidden1
Hidden2
Hidden3
Hidden4
Sum
Relative
importance
X1
0.0934
0.2669
0.0203
0.3347
0.72
17.88%
X2
0.7184
0.3795
0.7728
0.4192
2.29
57.25%
X3
0.1882
0.3535
0.2069
0.2462
0.99
24.87%
4.00
100%
Rank
3
1
2
7- Partial derivatives (PD):
To obtain the relative importance of each input, the partial derivatives of the ANN output with respect to
the inputs were computed (Dimopoulos et al., 1999). In a network with ni inputs, one hidden layer with nh
5698
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
neurons, and one output neuron, the partial derivatives of the output yj with respect to input xj and N total
number of observations are:
where Sj is the derivative of the output neuron with respect to its input, Ihj is the response of the hth hidden
neuron, who and wih are the connection weights between the output neuron and hth hidden neuron, and between
the ith input neuron and the hth hidden neuron. If the partial derivative is negative the output variable will tend to
decrease while the input variable increases. Inversely, if the partial derivatives are positive, the output variable
will tend to increase while the input variable also increases. The relative importance of the ANN output is
calculated by a sum of the square partial derivatives, SSD, obtained per input variable as follows:
The input variable which has the highest SSD value is the most important variable, and inversely the input
variable that has the lowest SSD value is the lowest important variable. The results of partial derivatives
methods are shown in Table (12).
Table 12: Relative importance and rank of inputs.
Partial derivatives
Inputs
X1
X2
X3
Total
Relative importance
0.16%
98.23%
1.61%
Rank
3
1
2
Comparison of the 7 methods used for quantifying relative importance of inputs:
Based on the results of Multiple Linear Regression (MLR), the Most Squares method was found to exhibit
the best overall performance compared to the other methods with regard to its accuracy, where the degree of
dissimilarity coefficient between MLR and Most Squares (MS) was the lowest, 0.2242, Table (13), both were in
the same cluster, Figure 4 and produced comparable results, Figure 3. The performance of Partial Derivatives
(PD), (Gevrey et al., 2003), was less than MS where the degree of dissimilarity coefficient between PD and
MLR was 0.7594. Dominance Analysis (DA) and Modified Connection Weights (MCW) showed moderate
performance where dissimilarities were 1.0842 and 1.8490, respectively. Connection Weights (CW) and
Garson’s Algorithm (GA) performed poorly where dissimilarities were 2.6719 and 4.1282, respectively. MS
method was found to accurately quantifying the relative importance of input variables and may be favored over
the other methods tested in this study. Such approach could successfully identify the relative importance of all
input variables in the neural network, including variables that exhibit both strong and weak partial correlations
with the dependent variable.
Table 13: Distance matrix based on Euclidian dissimilarity coefficient for the 7 relative importance methods.
5699
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
Fig. 3: Range plot showing relative importance of the inputs according to each of the 7 methods.
Fig. 4: Dendrogram showing cluster analysis (Ward method) of the 7 relative importance methods.
Conclusion:
The present study provides a comparison of the performance of 7 different methods for assessing variable
relative importance. Two algorithms are proposed to estimate the relative importance of the independent
variables. The performance of the proposed algorithms is studied using empirical data sets. Most Squares (MS)
is seemed to perform much better than the other methods, and agree with the results of multiple linear regression
in terms of the partial R2 and seem to be more reliable. The application of such effective models could
effectively offer better understanding of the variables govern specific crop parameter.
5700
J. Appl. Sci. Res., 9(11): 5692-5700, 2013
References
Afifi, A.A. and V. Clarke, 1990. Computer aided multivariate analysis, 2nd ed., Van Nostrand Reinhold, New
York.
Azen, R. and D.V. Budescu, 2003. The dominance analysis approach for comparing predictors in multiple
regression, Psychological Methods, 8: 129-148.
Budescu, D.V., 1993. Dominance analysis: A new approach to the problem of relative importance of predictors
in multiple regression. Psychological Bulletin, 114: 542-551.
Chapman, H.O. and P.E. Pratt, 1978. Methods of analysis for soil, plants and waters. Univ. Calif. Div. Agic. Sci.
Dimopoulos, I., J. Chronopoulos, A. Chronopoulou Sereli and S. Lek, 1999. Neural network models to study
relationships between lead concentration in grasses and permanent urban descriptors in Athens city
(Greece). Ecological Modelling, 120: 157-165.
Duh, M.S., A.M. Walker and J.Z. Ayanian, 1998. Epidemiologic interpretation of artificial neural networks,
American Journal of Epidemiology, 147: 1112-1122.
Garson, G.D., 1991. Interpreting neural network connection weights. Artif. Intell. Expert, 6: 47-51.
Gevrey, M., I. Dimopoulos and S. Lek, 2003. Review and comparison of methods to study the contribution of
variables in artificial neural network models. Ecol. Model., 160: 249-264.
Johnson, J.W., 2000 A heuristic methods for estimating the relative weight of predictor variables in multiple
regression, Multivariate Behavioral Research, 35: 1-19.
Khodary, S.E.A., 2004. Effect of salicylic acid on the growth, photosynthesis and carbohydrate metabolism in
salt-stressed maize plants. Intl. J. Agri. Biol., 6: 5-8.
Lebreton, J.M., R.E. Ployhart and R.T. Ladd, 2004. A Monte Carlo comparison of relative importance
methodologies, Organizational Research Methods, 7(3): 258-282.
Olden, J.D. and D.A. Jackson, 2002. Illuminating the “black box”: a randomization approach for understanding
variable contributions in artificial neural networks, Ecological Modeling, 154: 135-150.
Olden, J.D. M.K. Joy and R.G. Death, 2004. An accurate comparison of methods for quantifying variable
importance in artificial neural networks using simulated data, Ecological Modeling, 178: 389-397.
Özesmi, S.L. and U. Özesmi, 1999. Artificial neural network approaches to spatial habitat modeling with inter
specific interaction. Ecol. Model, 116: 15-31.
Ripley, B.D., 1996. Pattern recognition and neural networks. Cambridge University Press.
SAS Institute Inc., 2003. SAS Software v.9.1.3 sp4, Cary, NC: SAS Institute Inc.
Snedecor, G.W. and W.G. Cochran, 1982. Statistical Methods 7th ed., Iowa state Press Iowa, USA.
Tomassone, R., E. Lesquoy and C. Miller, 1983. La re´gression,Nouveaux Regards surune ancienne me´thode
Statistique. INRA, Paris.
Watts, M.J. and S.P. Worner, 2008. Using artificial neural networks to determine the relative contribution of
abiotic factors influencing the establishment of insect pest species, Ecological Informatics, 3: 64-74.
Weisberg, S., 1980. Applied Linear Regression. Wiley, New York.